“The world model not only predicts the future; robots are beginning to compete for the ability to “intervene in the world”.”
2026-09-07 08:00Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
5mentions
SummaryAI generated
In 2026, Guangxiang Technology, in collaboration with the research group led by Professor Li Shengbo from Tsinghua University, released the physical native world model Phi-WM 1.0 ActEffect. This model aims to meet the evolving needs of robots from “predicting the future” to “understanding the consequences of actions”. It features state decoupling representation, temporal causal driving, and physical law constraints, and incorporates action causality into the training process. Through distillation strategies, it reduces the computational cost of reasoning. In the LIBERO, LIBERO-PLUS, and RoboCasa-GR1 benchmark tests, ActEffect achieved average success rates of 98.8%, 80.3%, and 67.5%, respectively. Guangxiang Technology plans to integrate the world model, data, algorithms, and robot bodies into a single cycle, with the goal of compressing real-world training data to less than dozens of hours, in order to address the challenges of industrial scenario fragmentation and enhance generalization capabilities.
In 2026, the world model industry rapidly developed in the field of embodied intelligence, with many manufacturers adopting it as a core technology direction. Guangxiang Technology, in collaboration with the research group led by Professor Li Shengbo from Tsinghua University, released the physical-native world model Phi-WM 1.0 ActEffect, aimed at addressing the need for robots to upgrade from “predicting the future” to “understanding the consequences of actions”. This model features state decoupling representation, temporal causal driving, and physical law constraints, integrating action causality into the training process and reducing inference computational costs through distillation strategies. In the LIBERO, LIBERO-PLUS, and RoboCasa-GR1 benchmark tests, ActEffect achieved average success rates of 98.8%, 80.3%, and 67.5%, respectively. Guangxiang Technology plans to integrate world models, data, algorithms, and robot bodies into a single cycle, aiming to compress real-task training data to less than dozens of hours, in order to address the challenges of industrial scenario fragmentation and enhance generalization ability.